LAUSR

ABSTRACT This paper introduces a technique for reducing the sampling size necessary to explore a design space via building performance simulation. The strategy is based on a novel algorithmic procedure for creating a metamodel from an incomplete data series of a multivariate design problem. The algorithm interpolates the near-neighbours of the unknown vectors in a design space by the means of the near-neighbouring gradients, and progressively creates a global network of local relations involving vectors and gradients. The procedure produces low deviations with respect to the fully simulated series and performs at its best with sparse clusters of adjacent vectors (of the kind produced by star samplings or block-coordinate searches), unsmooth performance landscapes, and small sampling densities. The technique can also be utilized advantageously in optimization: post-metamodelling on the basis of the samples taken in an optimization search can indeed increase both the breadth of the information produced and the optimization efficiency.